Table of Contents
List of Figures
List of Tables
1.2. Problem Definition
1.3. Research Objectives
1.4. Research Methodology
1.4.1. Research Strategy
1.4.2. Data Collection
1.6. Thesis Outline
2. Literature Review
2.1. Introduction to Literature Review
2.2. Defining Artificial Intelligence
3. Applications of Artificial Intelligence
3.1. Customer Segmentation Using Self Organizing Feature Maps
3.2. Sales Forecasting Using Multilayered feed-forward Neural Networks
3.3. Introducing a Fuzzy Logic Expert System in International Product Pricing
4.Findings, Conclusion and Final Discussion
5. References and Bibliography
6.1. Appendix 1: Advantages and Disadvantages of Neural Networks
6.2. Appendix 2: Comparison of Neural Networks and Statistical Terminology
6.3. Appendix 3: Sample Questionnaire Used by I-Lance Ltd in Their B2B Market Research
6.4. Appendix 4: ARIMA Model Output in Gretl
6.5. Appendix 5: ARIMA Model Output in Gretl
6.6. Appendix 6: Holt-Winters Model Output in Eviews
6.7. Appendix 7: Holt-Winters Model Output in Eviews
6.8. Appendix 8: Five Steps in Implementing a Fuzzy Expert System in Matlab
LIST OF FIGURES
Figure 1 Thesis Research Framework
Figure 2 Six steps in market segmentation, targeting and positioning
Figure 3 Six steps in designing and implementing a SOFM for customer segmentation
Figure 4 The network of the segmentation study
Figure 5.1 Topological map of resulting market segments
Figure 5.2 Profiles of customer segments
Figure 6 A multilayered feed forward neural network
Figure 7 A sample architecture of a sales forecasting network
Figure 8 Six steps in designing a MFNN
Figure 9 Sales and price movements for motor gasoline sales
Figure 10 Correlogram of sales data
Figure 11 Architecture of a pricing expert system
Figure 12 Graph showing membership functions for fuzzy variable competition..
Figure 13 Fuzzy expert system
Figure 14 Five steps approach to constructing a fuzzy logic expert system
Figure 15 Determining the fuzzy variables to be used
Figure 16 Defining the fuzzy sets and membership functions in Matlab
Figure 17 Defining the case study rules in Matlab
Figure 18 Rule viewer in Matlab
LIST OF TABLES
Table 1.1: Profiles of resulting segments
Table 1.2: Profiles of resulting segments
Table 2: Common parameters in designing a back propagation neural network
Table 3: Summary results of statistical models with preprocessed data
Table 4: Summary results of statistical models with raw data
Table 5: Summary results of neural network models
This chapter introduces the concept of Artificial intelligence, its background and why it is an interesting field to research. The problem definition and objectives define the goals and intentions of the thesis while the research methodology serves as the basis of conducting a solid research study. Finally, the limitations are presented along with the outline of the thesis.
The exponential growth of information has motivated companies to gather and process information in a systematic way. This in turn has led to changes within companies in many directions. From the operational point of view, decision making procedures and mid-level managers are under constant pressure of catching up with new technologies and methods. From the strategic point of view we are witnessing a dramatic shift from the focus of creating competitive advantages through tangible and intangible assets to as Davenport (2007: 1) states “competing on analytics”.
Today’s business professionals are doomed to make their decisions in complex and data-rich environments. The use of analytical and quantitative methods is not regarded as the skill of scientists and researchers anymore. Job descriptions require knowledge not only of analytical methods but also as software packages that until recently were used only in laboratories and research institutions.
This trend is translated into the world of marketing by the marketing engineering approach. This approach emphasizes the need to use analytical methods in order to create efficient decision making in the marketing departments of companies. However, the focus lies more on using a predefined number of methods and attaining the skills to use them. We will not argue that this is a flaw. However, as still higher management emphasizes the need to educate managers and teach these methods and skills, nothing is done in the direction of improving the methods that managers are educated on.
Especially in marketing, these methods are based mostly on statistical and econometric techniques. These techniques have been around since a long time and later adapted to suit the needs of the marketing professional. However, their extensive use could not be a fact before the explosive growth of information technologies (IT) in the last decade. Apart from providing the basis for improving the analytical capabilities of companies, IT have created a growing stream of alternative analytical methods that compete with the current statistical ones.
These alternative methods fall under the broad category of artificial intelligence. According to Coppin (2004: 4) artificial intelligence involves using methods based on the intellectual behavior of humans and animals to solve complex problems. The roots of artificial intelligence date back to philosophers as Aristotle and Socrates. As Coppin (2004: 10) states, Plato wrote that his teacher Socrates intended to create an algorithm of describing the behavior of people and judging whether it was good or bad. Artificial intelligence continued to evolve and today represents a field of growing importance in academia and industry. Artificial intelligence covers a broad scope of methods and mainly Artificial Neural Networks, Fuzzy Logic and Genetic Algorithms. Their applications have been widely used in the production industry and in the past decade especially in the field of finance. It is not since quite a while that they have entered other business related spheres and the number of publications point out in this direction.
However, the main criticism of these methods has been their highly computational complexity often difficultly understood by marketing professionals as compared to industrial engineers and finance “quants”. Their widespread use though in marketing, meaningless to mention international marketing has not yet been seen because of the above-mentioned factors. Additionally no accessible materials and publications especially for this target group exist.
The need to cover the topic in an accessible and practical way as well as concentrating in the field of international marketing arises. In the global labor market where today’s professionals need to be equipped with qualitative and quantitative skills, the ability to master new and alternative methods is one of the most important requirements. Exploring the capabilities of AI in international marketing will be a challenge for everyone.
The recent growth of interest of international marketing professionals in artificial intelligence methods is linked with the nature of the information with which they operate. The information available to the international marketing manager is highly heterogeneous, containing insights about consumer preferences or sales prices that have a high level of complexity because of the international element inherent in the data. These elements may include for example cultural influences on the preferences of consumers, regulatory differences, taxes and tariffs on prices or different structures of the distribution channels. Apart from having to account for the influence of the above mentioned elements not present in the local market, the marketing manager has to acknowledge that the relationships in international marketing data are non linear in their nature and not normally distributed, making it impossible for classical statistical models to handle them. Exactly this complexity of the international environment and information available to the managers opens new opportunities for applying artificial intelligence methods in international marketing.
This thesis adds value to current research in a number of ways, mainly linked to the target audience’s interests. Within this audience, we have identified two target groups. Firstly, this research targets marketing professionals interested in adopting an interdisciplinary approach to the study of international marketing problems as well as other marketing practitioners who are seeking for new perspectives on how to build models for everyday operations. This research will provide the algorithms of how, different artificial intelligence methods are applied depending on the context and providing comparisons with already existing statistical models. Secondly, academic researchers would be interested in getting a more in-depth look at the applications of artificial intelligence in international marketing and the evaluation of these techniques based on indicators that are important not only for the reliability of a model but as well as on indicators relevant to business professionals. After all theories and models are built into laboratories, however they still have to pass the reality test, to be used successfully by practitioners.
1.2 Problem Definition
As already discussed, artificial intelligence methods are becoming more and more popular for today’s business professionals. Throughout our literature review, we identified a main gap that lies in the very academic examination of the topic, which in most cases is directed towards researchers and not practitioners. Furthermore, not much attention has been given to the actual application of Artificial intelligence in international marketing. A natural question that emerges is to what extent and how can these methods be applied in international marketing by business professionals.
In order to address this gap identified through the literature review, we have set the aim of exploring artificial intelligence applications in international marketing as a tool for achieving competitive advantage, as well as how and to what extent these methods can be useful to practitioners.
In pursuing the aim of this thesis we will present thoroughly a selection of the most widely used methods of artificial intelligence. These will include neural networks and fuzzy logic. At the same time, we shall cover the underlying theoretical foundations of forecasting methods, customer segmentation and expert systems in order to present the different applications more successfully.
1.3 Research Objectives
In trying to achieve its main aim this study has set the following objectives:
1 Identify the main advantages and disadvantages linked with the use of artificial intelligence methods in international marketing
2 Evaluate critically the applications of artificial intelligence methods in international marketing and compare them to alternative statistical ones
3 Explore the whole process of implementing artificial intelligence methods from a holistic perspective, thus covering context and non-context related factors, and additionally accounting for the international perspective of marketing.
4 Formulate recommendations for further research and algorithms of how to practically employ artificial intelligence methods in solving tasks in international marketing.
It would be a mistake for the reader to view each of the stated research objectives as separate, unrelated activities. The answers to these objectives will consequently provide information for the achievement of the research aim which is to gain insight into the efficiency and practical implementation of artificial intelligence methods in international marketing.
1.4 Research Methodology
In this section, we will discuss and justify the research strategy along with the presentation of data collection techniques adopted in the empirical chapter of the thesis. Details on the sampling methods are provided together with a discussion on quality standards and the notions of reliability and validity of the thesis.
1.4.1 Research Strategy
The research philosophy adopted by this researcher and thus used throughout this study is positivism. Positivism emerged as a philosophical approach in research methodology in the 20th century and ideally suits a quantitative study of such a manner. However, as we will explain later in this chapter, deviations from some of the principles or axioms of positivism as a scientific method are allowed as our thesis represents an interdisciplinary study covering topics from both the social and natural sciences. The use of one method alone would not be possible.
One of these deviations is the principle of falsification proposed by Karl Popper and concerns the very beginning of the research study. As McNeil (2005: 178) notes, this principle argues that instead of looking for evidence to prove a hypothesis right, scientists should look for evidence that proves it false .This approach requires that scientists are rigorously skeptical and suspicious of their own hypotheses. The approach of being critical and skeptical, and ready to reject the theories and hypotheses that we fundamentally believe in, when evidence are against them, will prove to be the key to producing reliable findings worth noticing.
Following the work of positivist methodology, we have used a deductive approach throughout our research. We have reviewed different theoretical concepts from the fields of international marketing, and more precisely sales, pricing strategies and customer segmentation. Additionally we have reviewed the statistical and artificial intelligence methodology needed to be applied in the empirical section of the thesis. Based on these theoretical foundations we have tried to test hypotheses associated with our objectives. Furthermore, at the end of the thesis we try to generalize our findings based on our empirical results.
After choosing the methodology of approaching a research topic, the researcher should choose a research design. The research design is on one hand determined by the researcher’s philosophical positioning as discussed previously and on the other hand by the nature, aims and context of the research. For the purposes of this study the explanatory and descriptive approaches are used.
The use of two methods rather than one does not represent any nuance for the research community. As McNeill (2005: 8) states, “any explanation requires description, and it is difficult, or perhaps impossible, to describe something without at the same time explaining it”. The descriptive approach is needed as we seek to gain an understanding on existing knowledge and research on artificial intelligence applications in international marketing. Moreover, the fact that this thesis covers broad and diverse topics with the application of cases, it will be evident that descriptive and explanatory methods are used interchangeably throughout the different chapters.
As mentioned earlier deviations from positivist principles are allowed in this study and this concerns the use of exploratory research as an additional research strategy. Exploratory research corresponds to our intention of bringing new knowledge in the field of AI applications in international marketing. Exploratory research is applied through the case study method.
Case studies represent a powerful tool when embarking into management research and their use nowadays is expected as an informal requirement. In order to give answers to part of our objectives, we present three distinct cases . The case method provides us with the opportunity to implement a profound analysis on decision making methods in international marketing, focusing on a specific unit of analysis the marketing departments of companies. At the same time we are able to account for the complexity of the environment, as decision makers and companies are placed in a broader context, the market segment in which they operate. For illustration, let’s say that we have covered the objectives of comparing the effectiveness of artificial intelligence methods to statistical ones but at the same time considered the context of the decision maker and outlined some additional metrics of comparison apart from standard metrics that tend to be global and applied for all industries. The use of the case method, however, is not without its limitations that the researcher is aware of. Case studies limit the ability of the research to reach any generalizations (Biggam 2008: 123).
However, this particular study intends to add knowledge to the research concerning artificial intelligence uses in marketing and the case study method will provide us with additional and specific observations on the micro-level.
As mentioned earlier this study represents an interdisciplinary research. Throughout this section it was made evident that a combination of research strategies will be used based on the positivist scientific method. Below our research framework is presented. In the next section, we will present the data collection methods in more detail.
illustration not visible in this excerpt
1.4.2 Data Collection
Conducting a research study involves the process of data collection. Data collection facilitates the researcher with the quantitative and qualitative information needed to carry out his research successfully. As Stolley (2005: 209) states, professional researchers tend to mix and match qualitative and quantitative research strategies, which has its impact on the data collection methods. Because of this reason and the fact that the nature of this research is interdisciplinary, we have collected both qualitative and quantitative data. We have additionally applied the method of triangulation in order to ensure the reliability of the collected data.
We will collect qualitative data mainly based on interviews. As Biggam (2008: 83) notes, interviews are an appropriate means of collecting qualitative data, and commonly used in case studies. Semi-structured interviews with senior management will provide data concerning the pricing case study. Structured interviews and filling out questionnaires will be the main source of data for the customer segmentation case study. In the course of this particular case study, the researcher will take part in some of these interviews in order to ensure and control the reliability of the provided data.
Secondary data will be retrieved from company databases and will be used for the sales forecasting exercise. Additional secondary data will be gathered from company materials and reports, industry and academia publications, as well as observations made by the researcher at the companies’ premises.
This thesis takes on the discussion of applying artificial intelligence applications in international marketing. The scope of the topic and the concrete applications are too numerous to be covered here. We have presented a representative number of applications of neural networks and fuzzy logic. Genetic algorithms and partially ANN for classification are only vaguely mentioned. We acknowledge the existence of different perspectives on the main theme of the thesis as well as within parts of the thesis from researchers in different industries. However, we feel confident that this thesis has accounted for these differences and presents extensive and critical results based on the objectives set in the beginning.
Limitations exist in the data collection procedure as well. Data and case companies chosen have been solely selected based on convenience sampling and mainly willingness to provide data and assistance. This has lowered our flexibility of choosing on one side global companies who would better suit our profile giving more international relevance on the thesis and on the other side receiving more detailed data.
Limitations of time and resources are classical constraints in research work and this thesis is not an exception. Having the ability for example to selectively present a greater number of companies would give the thesis more gravity and relevance. Last but not least, sharing knowledge with professionals and academics from six nations has been difficult and time consuming but at the same time provided the researcher with key knowledge, and raised the expectations of what should be attained by this thesis.
1.6 Thesis Outline
The first chapter presents the background of the subject of this thesis. Furthermore, the aim and objectives of the research are provided. Then the research methodology is presented in detail. An emphasis is given to the research design and data collection methods in order to gain further insight on how the study was conducted and its methodological foundations. Finally yet importantly, the limitations of the study along with the thesis outline are presented.
2. Literature review
The literature review chapter presents the process of compiling literature concerning most of the topics dealt with in this thesis. First, a general survey of business applications of Artificial intelligence is provided. The review focuses on AI applications for marketing and namely consumer segmentation, sales forecasting and new product pricing expert systems. Secondly, some literature on forecasting and market segmentation theories, relevant to the thesis, is briefly summarized.
3. Artificial intelligence applications
This chapter is constructed on the case study methodology and consists of three independent applications of artificial intelligence in international marketing. Each case is structured in a similar fashion in order to make it easier to follow the discussion on the applications. The case studies’ content includes the theoretical background of the problem solved from an IT and a marketing perspective, an analysis of the solution with recommendations for the company’s management and a section of findings.
4. Findings, Conclusions and Final Discussion
The last chapter is divided into three parts. The findings section presents the main results obtained in the empirical chapter of the thesis. In the following part conclusions from the theoretical and empirical chapters are drawn. Moreover, in this section, the thesis is linked back to the aims and objectives and the research questions are answered. In the section presenting the final discussion, other perspectives not presented in the thesis are covered, as well as suggestions for further research are outlined.
2. LITERATURE REVIEW
This chapter will examine the main issues surrounding the applications of artificial intelligence as presented in selected publications and textbooks, how these methods could be useful to practitioners and ultimately give answers to part of our research objectives.
2.1 Introduction to Literature Review
The literature review will focus on the first three objectives set out in the introductory chapter of this thesis (the final objective – objective 4 – is derived from results of the empirical section of thesis as well as the findings from objectives 1, 2 and 3). By analyzing and then synthesizing the areas of literature concerned with our objectives, we aim at identifying and discussing the gaps of current research which need to be covered by this thesis.
Firstly the main advantages and disadvantages linked with the use of artificial intelligence methods in international marketing will be examined. Then these applications will be critically evaluated and compared to traditional statistical methods. Additionally, the whole process of applying these methods will be assessed from the practitioners’ point of view and much weight will be given to context and non-context related factors.
To conclude with, the value of reviewing the aforementioned literature areas will be to provide an analysis of the thesis topic in a structured way, and to facilitate us with the skills and insights needed to apply these methods in the empirical section of the thesis. Moreover, we hope that at the end of this section, a clear justification of the research objectives set will emerge and additionally the reader will be acquainted with the main issues surrounding artificial intelligence in international marketing.
2.2 Defining Artificial Intelligence
A sensible starting point to the literature review would be to investigate what artificial intelligence represents. There has been much debate on the proper definition of artificial intelligence and different authors give more emphasis on different qualities of artificial intelligence. We believe that Engelbrecht (2007: 30) defines artificial intelligence quite properly. Artificial intelligence represents a combination of several research disciplines as computer science, physiology, philosophy, sociology and biology that led to the modeling of biological and natural intelligence creating the so-called “intelligent systems” or algorithms. According to Binner et al. (2004: 1) these algorithms represent data driven methodologies designed to solve real world complex problems and include Artificial Neural Networks (ANN), Genetic algorithms (GA), fuzzy logic, probabilistic belief networks and machine learning. A more simple definition is that of the IEEE Neural Networks Council of 1996: the “study of how to make computers do things at which people are doing better”. In this thesis, we will discuss the first two methods: ANN and fuzzy logic, which we define below.
Artificial Neural Networks
According to Smith and Gupta (2002: 1) “ANN’s are simple computational tools for examining data and developing models that help identify interesting patterns or structures in the data”. ANN can learn to perform the following tasks by only observing patterns and characteristics in the training data provided by the researcher:
1. Predict future events (Forecasting)
2. Classify unseen data into predefined groups (Classification)
3. Cluster provided data into homogenous groups (Clustering)
According to Boyadjiev and Boyadjiev (2007: xv) fuzzy logic is a form of logic that deals with objects that are a matter of degree, with all possible values of truth between “yes” and “no”, as opposed to classical logic that distinguishes between “yes” and “no”. Similarly, Cox (2005: 160) describes fuzzy logic as the “logic of continuous variables” and classical logic as the “logic of discrete variables”. In this sense fuzzy logic can quantify and present the opinions of experts expressed with the vagueness of natural language (Krichevskyi 2005: 74). Fuzzy logic is able to capture expressions like: the “water is warm” or “the water is cold”, “John is a tall person” and many others. Even if fuzzy logic sounds extremely strange and is regarded as not popular, its applications are present in our everyday life. These are applications in washing machines, cars, elevator control mechanisms and many others.
As far as international marketing applications are concerned, fuzzy logic is able to deal with the ambiguity inherent in marketing decision making. By being able to incorporate linguistic variables as “very high”, “low” or “medium”, as opposed to traditional expert systems, an expert system based on fuzzy logic can assist marketing managers in making pricing decisions, crafting strategies, implementing SWOT analyses and others. Such a system could furthermore facilitate the standardization of decision making procedures within the subsidiaries of a global company. However, we will discuss the capabilities of fuzzy logic linked to business and marketing in more depth later. We continue our literature review with the sections dedicated on the findings concerning with our research objectives.
2.2.1 Literature Review of Research Objective 1
“Identify the Main Advantages and Disadvantages Linked with the Use of Artificial Intelligence Methods in International Marketing”
184.108.40.206 Advantages and disadvantages of ANN’s
In this section, we identify the advantages and disadvantages of ANN and fuzzy logic in two distinct sub-sections.
The advantages and disadvantages concerning the applications of ANN are present in almost every study carried out in this field. Below we present thoroughly the most important advantages and disadvantages cited in the studies that we have reviewed as well as methods to tackle the shortcomings of ANN.
Vellido et al. (1999) have reviewed more than 100 research papers from 1992-1998 and have come up with a quite detailed description of the advantages and disadvantages of ANN’s. We have used the aforementioned study, the findings of Cox (2005: 53) and the advantages and disadvantages identified by our own research. We have ranked the most important ones according to the work of Vellido et al. (1999) and our own estimates, in ascending order.
- ANN’s are able to learn any complex non-linear mapping and are generally regarded as universal function approximators. Thus ANN’s learn from “data or experience” and perform well in “data rich-knowledge poor situations”.
- As non-parametric methods, ANN’s do not make a priori assumpions about the distribution of the data and they are very flexible with missing,incomplete and noisy data (Venugopal, Baets 1994: 18). Thus, they are less susceptible to model misspecification (bias) and statistical tests shortcomings, than parametric methods.
- ANN can be easily updated and suitable for dynamic environments
- ANN can be used efficiently in forecasting as they can find non-linear structures in a problem and at the same time are able to model linear processes (Zhang 2002: 2)
- Hidden nodes in MFNN can be regarded as latent,unobservable variables and thus used efficiently in market response models by modeling the unobserved processes in the mind of the consumer (Parsons and Dixit 2002: 26)
- The main shortcoming of ANN’s according to Vellido et al. (1999 :62) is their nature of black boxes. This means that there is no concrete answer to the behavior of the network and to how the network has reached the presented solution. For the same reason, a solution provided by an ANN is neither easily understood nor validated because information processed in the network is a set of weights and connections that provides no insight as to how a task is performed. That is, there is no clear relationship between the input and the output.
- No formal rules of techniques for the use of ANN. Non linear methods still lack rules of how to assess the relative relevance of independent variables. Consequently, the selection of variables is based mainly on techniques used with linear methods.
- Overfitting has been cited by many authors as one of the major drawbacks of ANN’s. According to Cox (2005: 33), ANN can overfit training data becoming useless in terms of generalization. “Overfitting occurs when the network is very accurate with the training data used to build the network, however performs very poorly when presented with new data”.(Zhang 2002: 7)
- The selection of the network topology and its parameters lacks theoretical background. It is still a trial and error procedure. As Kaastra and Boyd (1996: 216) mention especially for the forecasting applications of ANN, the large numbers of parameters that must be experimentally selected to generate a good forecast is considered as a big drawback for ANN’s.
- The lack of a formal background in optimising the neural network architectures comes second among quoted drawbacks in the study of Vellido et al. (1999: 65)
- Cox (2005: 53) identifies several practical disadvantages like the need for deep expertise and knowledge in neural network design in order to be able to construct an efficient network. Additionally it requires a large amount of data and can handle non-numeric data very difficultly.
Most of the disadvantages mentioned can be overcome with the use of other AI methods together with ANN’s. The use of rule based fuzzy systems (Benitez et al. 1996) for example eliminates some of the drawbacks of the black box methodology. However, we need to comment that fuzzy systems can be combined with ANN to solve only a limited scope of business related problems. Genetic algorithms also have helped in this direction insofar they can be used to avoid the problem of overfitting and choosing the best architecture for the network. The increased capabilities of computers have also contributed to the faster and more convenient use of ANN’s. Ultimately, as Kaastra and Boyd (1996: 216) mention, the critical success factors of applying ANN lie in the ability of the researcher to have patience, time and resources to experiment. Nevertheless, applying ANN’s remains to some extent a heuristic procedure that involves in some degree experimentation and the lack of formal rules surrounding their application real business problems.
220.127.116.11. Advantages and disadvantages of Fuzzy logic
Hutcheson and Moutinho (2008: 204), Cox (2005: 55) and Krichevskyi (2005: 77) identify several advantages and disadvantages of using fuzzy logic and fuzzy sets the most important of which:
- Very easy method to apply and understand both because is based on the natural language of communication as well as because it is based on solid mathematical foundations
- Takes into account the skills and knowledge of experts
- Can model non-linear functions
- The membership function is designed so as to treat the vagueness caused by natural language. Therefore is more accurate than statistical decision making models
- The membership function standardizes the semantic meaning variables and makes the method easily applicable in different environments, markets and consumers
- Fuzziness encompassing the uncertainty in decision making makes the models closer to the real uncertain business environment
- “Semantics tailored to specific cognitive model (e.g., “young” in marketing is likely not the same as “young” in the credit department)”
- Difficult measurement,scaling and estimation of the values of the parameters of the model
- Few knowledge engineers trained in fuzzy logic (and those that do have some fuzzy logic experience generally have little if any experience with database systems and business problems)
- Fuzzy logic expert systems cannot discover deep relationships
- Successful applications depend on the proper definition of fuzzy sets
- Applicable to numeric data only
There are also disadvantages that come from a philosophical view. According to this point of view, fuzzy logic simply cannot co-exist with classical logic. This is however a fundamentally methodological argument which we will not analyze further. What we are interested in are the mentioned disadvantages. Those identified disadvantages we acknowledge and much of the empirical work that will be carried out will be in the direction of presenting ways of how to eliminate those disadvantages. In our opinion, the above-mentioned disadvantages are mainly based on the lack of an in-depth analysis of case studies.
2.2.2 Literature Review of Research Objective 2
“Evaluate Critically the Applications of Artificial Intelligence Methods in International Marketing and Compare Them to Alternative Statistical Ones”
The goal of the literature review concerned with the second objective is to evaluate the applications of artificial intelligence and present the benchmarks against which AI methods are compared. The applications of artificial intelligence are present in a wide range of business related problems. Extensive reviews in this field especially, have not been undertaken recently and those that exist focus mainly on the applications of neural networks. That is why we have analyzed ANN’s applications separately from fuzzy logic and done that in a time consuming manner, insofar that we needed to review a great deal of up to date case studies with applications that we were interested in.
The most comprehensive studies are those of Smith and Gupta (2000), and Vellido et al. (1998). They review applications of ANN’s ranging from insurance, taxation and finance to real estate prices forecasting and marketing. From those studies we only concentrate on the sections that survey methods applied in marketing and management. We also consider two textbooks, Smith and Gupta (2002) and Zhang (2004) that present more than 35 case studies with applications of ANN in business.
In the field of marketing, market segmentation studies and those related to retail cover much of the proportion of research done in this direction. More precisely these studies include applications with market response models, clustering analysis, classification, market share forecasting and retail sales forecasting. According to the cases presented in Smith and Gupta (2002), the applications are divided based on the particular network used.
Those we are interested mainly in are Self-organizing feature maps (SOFM) and Multi Feed-forward Neural Networks (MFNN). SOFM’s are applied to clustering applications in a variety of marketing problems like market segmentation, brand and customer behavior analysis. MFNN are used for classification applications and solve problems related to target marketing, customer retention and loyalty decisions. However, it is already made apparent that the main field of MFNN applications is prediction. The above-mentioned authors present an abundant number of case studies of MFNN applied in sales forecasting, market development forecasting, brand choice models (Parson 2002: 36) as well as customer response forecasting. We have considered all of these studies in more detail especially for answering objective 3 and to some extent objective 1.
In order to evaluate the applications of AI in marketing we needed first to identify them, as described in the preceding paragraph. Then evaluating the effectiveness of the applications continued as a two-step process. Firstly, we identified the benchmarks against which ANN’s applications are measured. Below we present a detailed description of these benchmarks according to the method concerned. Secondly, we measured the effectiveness of ANN’s applications based on other measures that do not concentrate on the model and the solutions’ results, rather on other indicators that are more important for practitioners.
18.104.22.168 Evaluating the forecasting applications with MFNN
We begin with the benchmark models of the forecasting exercises. The concrete forecasting application that we are interested in is sales forecasting. Sales time series have special properties like seasonal and trend components. That is why different models and measures are used to evaluate these specific applications rather than for example financial time series. Vellido et al.(1999) cover the architecture characteristics of the networks and what models are used as benchmarks against the performance of the neural networks. We have to specify that the base ANN used in all studies is the MFNN based on the backpropagation algorithm.
The benchmark models
In Zhang (2002: 55) sales forecasts based on ANN’s are benchmarked against mainly autoregressive integrated moving average (ARIMA) models. Remus and O Connor (2001: 247) compare forecasting models based on MFNN to Box-Jenkings techniques (ARIMA), deseasonalized exponential smoothing, judgmental methods and a combination of the above mentioned. Frank et al. (2003) forecast women’s apparel sales using a wide variety of methods and compare the performance of the MFNN to those of the Holt-winters exponential smoothing technique, ARIMA, Fourier analysis and moving averages. The most comprehensive analysis of the above-mentioned forecasting methods is available in our opinion in Armstrong (2001).
All the presented models are compared based on their forecasting accuracy. The forecasting accuracy is measured with the Mean Squared Error (MSE), the Mean Absolute Percentage Error (MAPE) and the Mean Absolute deviation (MAD).
Additionally to the above-mentioned models, Zhang (2002: 13) proposes additional two criteria for evaluating the most accurate forecasting method. His recommendations have mainly to do with the process of conducting the forecast. Firstly, he proposes that true out of sample data have to be used in the evaluation procedure and secondly, enough sample size has to be ensured (40 for classification studies and 75 for time series).
22.214.171.124 Evaluating clustering applications with the use of SOFM’s:
Our literature review in this sub-section was focused on clustering and segmentation methods used especially in marketing applications. In the majority of the studies we have earlier mentioned the SOFM is benchmarked against the K-Means clustering algorithm and in some cases against linear discriminant analysis or the logistic regression.
Measuring the efficiency of the models
Cardoso and Moura-Pires (2002: 45) use a couple of methods to compare the efficiency of the clustering results between the k-means algorithm and the SOFM. These include ANOVA tests, analysis of the within group variance as well as the weighted absolute deviation of the clusters. Similarly Kiang et al. (2002: 154) measure the accuracy of the cluster assignments using the pooled within cluster variance or the sum of squares (SSE). There is also a number of studies that do not measure the efficiency of the SOFM method as it is only used as a method of variable reduction, similar to factor analysis. Subsequently, after the reduction of the dimension of the variables and the number of clusters is identified, the K-Means algorithm is used to create the clusters within the variables, thus the SOFM and the K-Means algorithm are used together in the solution of the segmentation problem and not competitively.
As we already said, the comparison of the different ANN models should satisfy the criteria of practitioners as well. Until now as we already saw academics put more weight in the performance of the model and have some unstructured views concerning the underlying statistical assumptions and data needs. The Fractal Analytics whitepaper on Classification techniques (2005) introduces a couple of additional evaluation indicators that are more qualitative but not the less relevant. These include a detailed and structured analysis of the already mentioned criteria as well as the criteria of complexity of deployment, transparency and model building time. Complexity of deployment is concerned with the complexity of creating and maintaining a neural network. Transparency measure the extent to which the network acts as a black box or it is constructed in such a way that its operations are more understandable. Finally, model building time is also something that every practitioner is concerned with. In the analytical section of the thesis we will try to summarize the models presented in the empirical section based on the above mentioned quantitative and qualitative criteria.
126.96.36.199 Fuzzy logic applications
Fuzzy logic is applied in business in a number of ways. As our main focus is to create a fuzzy expert system for decision making we will avoid commenting on its other applications. In Boyadjiev (2007) and Zopounidis et al. (2001) numerous case studies of fuzzy logic applications are presented. These include fuzzy decision making applications in marketing, market forecasting and market analysis applications, hybrid systems of fuzzy logic and neural networks used in market segmentation and others. According to Hutcheson and Moutinho (2008: 204), the uses of fuzzy logic in marketing include modeling consumer behavior, marketing planning (Diagnosis and Prognosis), new product testing, perceived price testing and marketing communication effects research.
The classical example of fuzzy decision making that concerns marketing and will be carried out empirically in the next section of this thesis, is the application of new product pricing. This application is presented in Cox (2005: 155) and quite extensively with a variety of examples in Boyadjiev (2007: 91).
The above-mentioned studies are quite useful for evaluating the applications of fuzzy logic. They cover extensively the mathematical sequence in which a solution is reached as well as how to construct the fuzzy expert system. However, there is a lack of a practical nature of the design of these applications, in the sense that specific questions that concern practitioners are not treated in detail. These are questions that deal with the specification of the membership functions, how to approach the experts and extract their knowledge and last but not least, the questions arising from the discussion around the pricing strategy formulation within companies and how it can be translated practically in the fuzzy expert system. What is also important for this thesis, is to analyze the application of such a system especially in international marketing and hopefully contribute to the knowledge and results of current research in the field and add to the work done in the direction of popularizing fuzzy logic applications in marketing and management. As Krichevskyi (2005: 140) mentions, the abilities of fuzzy methods in management are still undermined by current professionals and have not received the attention they deserve. We will try to give an answer to these problems with our empirical exercise performed in the next section.
2.2.3 Literature Review of Research Objective 3
“Explore the Whole Process of Implementing Artificial Intelligence Methods from a Holistic Perspective, thus Covering Context and Non–Context Related Factors, and Additionally Accounting for the International Perspective of Marketing”
Throughout our literature review we have reviewed more than 40 case studies concerning the field of AI. With some exemptions, all case studies are focused on the academic perspective of the presented problem. By the latter we understand the focus on applying the methods and measuring their accuracy and thus avoiding the discussion of issues related to the marketing problems solved. Authors forecast the sales of a product or segment a particular market and then discuss the accuracy of the solutions. No discussion arises concerning the needs of performing these methods, as well as what the implications of the results for the marketing professional are. This drawback exists because, although the data sets used in the majority of the case studies are from real companies, the problems that are solved do not come from business reality, thus they are not linked with the problems faced by the practitioner. This is a big disadvantage of all these papers as practitioners are interested in problems and questions which arise in the course of applying these methods in real world marketing problems. We assume these to be context and non-context factors that are related with the whole process of applying AI applications in solving international marketing problems. That is why this particular objective will be answered by applying AI methods based on real marketing problems in the empirical section of the thesis.